This Machine Broke CAPTCHA So We Can Have Cooking Robots

The artificial intelligence startup Vicarious announced a new program today that it claims is the closest a machine has come to being able to think, and more importantly, see like a human being.

The startup claims its AI software is able to crack most CAPTCHAs, the internet's gold standard for distinguishing between people and machines, between 90 to 95 percent of the time. That includes reCAPTCHAs, the new-and-improved version of the reverse Turing test that's widely used across the web, including by Google, Yahoo, and PayPal.

It’s no secret that CAPTCHAs aren't a panacea for filtering out bots and spammers; security researchers and password hackers have been cracking the codes for years, but usually it requires brute force.

Vicarious' so-called breakthrough isn't about security; it's about artificial intelligence. Like many other tech companies, it's working to develop "thinking" machines that mimic the human brain—specifically the neocortex, the most recently developed part of the brain that’s responsible for sensory awareness, perception, and consciousness.

But unlike the machine-learning systems behind Siri and Google Voice, the San Francisco startup is putting the focus on visual intelligence. It's trying to automate the eyes. By focusing on visual perception, and developing algorithms to mimic some of the brain's cognition that's not specific to this focus, the software is able to crack CAPTCHAs with intelligence instead of pure computational power, Vicarious claims.

Visual automation could have major implications for building humanoid robots, better medical imaging, image and video search, or anything else that combines machine-level intelligence with human-like sight—"like letting you know how many calories you're about to eat by looking at your lunch," MIT Technology Review wrote. Eventually, a computer could even possess the oh-so-human quality of imagination. It could see a dog in the clouds, Vicarious cofounder Dileep George said.

Breaking CAPTCHA wasn’t the goal, another Vicarious co-founder, Scott Phoenix, told Science. “It was just a sanity check. We believe that higher level intelligences are all built on the somatosensory system. So that’s why we started with vision.” Once connected to a robot, the next benchmark could be, for example, “preparing a meal in an arbitrary kitchen.”

Vicarious is calling its CAPTCHA experiment and the algorithm behind it—the Recursive Cortical Network™—a big win for the company. And if you're trying to create a computer that thinks like a human, what better way to test it than with the go-to method for telling computers and humans apart.

Still, while human discretionary powers prevail, the company's claims must be taken with some skepticism. Vicarious has released a video demonstration (below) of its AI software outsmarting CAPTCHAs, but it hasn't yet published the research paper describing the technology, so experts haven't had a chance to review the process and poke holes in it. That also means, however, that bad actors haven't been able to use the approach to exploit CAPTCHA and threaten the security of the entire internet.

The startup also says it's not planning to release the software commercially, and if they do it won't be to disrupt the internet's CAPTCHA system, but to build futurist applications like robots and medical imaging.

Meanwhile, Luis von Ahn, the computer scientist at Carnegie Mellon who created CAPTCHA, is not convinced. "This is the 50th time somebody claims this," he wrote to ScienceNOW. "I don't really get how they think this is news :) If their program is actually a break, we can simply add more distortion or switch to image-based CAPTCHAs."

That said, it's clear that traditional CAPTCHAs are nearing their end of their tenure. (I can't say I'm sorry—those things are a major pain.) Already security researchers are floating more secure alternatives to the tests.

Google announced just three days ago it's improving its reCAPTCHA to include more complex indicators like colors, 3D shapes, and lighting angles. Other researchers have floated the idea of Rorschach-style ink blot tests that prompt humans to interpret the image with a word or phrase and recall the association later. That’s a distinctively human skill beyond even the "smartest" computers. For now, at least.

Also see

Facebook Is Building an Artificially Intelligent Brain with Your Data

I/O Shows That Google's Ultimate Search Is for Artificial Intelligence

Let's All Now Form an Orderly Line and Order a Perfect Drink from the Robot Bartender